Nvidia Earth-2 Open Models Span the Whole Weather Stack (huggingface.co)

NVIDIA has unveiled three new open-source models under its Earth-2 initiative, designed to enhance weather forecasting applications across the full spectrum of meteorological tasks. This announcement is significant for the AI/ML community as it democratizes access to cutting-edge weather and climate models, allowing developers to create customized simulations tailored to their specific data and infrastructure needs. The models—Earth-2 Nowcasting, Earth-2 Medium Range, and Earth-2 Global Data Assimilation—are now available on Hugging Face and leverage advanced architectures like StormScope and Atlas, showcasing superior predictive capabilities compared to traditional physics-based models. Earth-2 Nowcasting delivers rapid, high-resolution storm forecasts within minutes by using generative AI to directly simulate storm dynamics, significantly outperforming existing models in short-term precipitation predictions. Earth-2 Medium Range forecasts extend up to 15 days by using a latent diffusion transformer architecture that effectively preserves atmospheric structures, while the upcoming Earth-2 Global Data Assimilation model promises rapid initial condition generation, streamlining the forecasting process from hours to seconds on GPU hardware. Collectively, these innovations not only enhance the accuracy of weather predictions but also empower developers to build sovereign and scalable climate solutions, marking a transformative step in the integration of AI within meteorological science.